Joshua Yeats

joshuayeats@anthropometric.co.uk •github.com/yeatsy

BSc Artificial Intelligence  •  University of Edinburgh, Class of 2025

Welcome. I'm a recent AI graduate from the University of Edinburgh with a passion for building data-driven products that transform complex signals into clear, actionable insights.

Below you'll find my honours dissertation on learning from sparse data, an iOS app for diabetes management, and various open-source ML tools. I'm now seeking graduate or internship roles where I can apply my skills in ML engineering and product design to solve meaningful problems.

Personal

My fascination with AI isn't just academic—it's personal. I'm driven by a desire to apply artificial intelligence to solve tangible, real-world problems, a passion that led me to create projects like Islet. As someone living with Type 1 diabetes, I have a firsthand understanding of the challenges of managing a chronic condition, which fuels my motivation to build tools that genuinely improve people's lives.

Beyond the world of code and algorithms, I'm an avid crossfiter, runner and cyclist. I find that the discipline and perseverance from sports translate directly into my work, pushing me to build more resilient and effective solutions.

• Honours Dissertation – Learning from Sparse Data

  Undergraduate thesis evaluating program synthesis, LLM rule generation, and RL for interpretable decision support in Type 1 diabetes care.

  PythonLLMReinforcement LearningProgram Synthesis
• Islet Diabetes

  iOS app for personal diabetes management – syncs Apple Health, tracks glucose, medication & meals, and uses GPT-4o mini for personalised insights.

  SwiftiOSGPT-4oApple Health
• LSTM_BGF

  PyTorch implementation of a Long Short-Term Memory network for blood-glucose forecasting.

  PythonPyTorchLSTM
• xLSTM_BGF

  Extended LSTM architecture with attention and residual connections for improved glucose prediction.

  PythonPyTorchLSTMAttention

Honours Dissertation

Learning from Sparse Data – TL;DR. This study compares three ways to derive transparent insulin-dosing rules from noisy, low-frequency CGM data: (1) statistical SyGuS program synthesis, (2) large-language-model prompt engineering for rule generation, and (3) Q-learning reinforcement learning. SyGuS rules were too coarse to capture physiological variability; LLM-generated rules reached high predictive accuracy while remaining human-readable; Q-learning produced adaptive policies keeping simulated glucose largely in-range. Results suggest a hybrid pipeline that seeds RL with LLM-derived priors is the most promising path toward interpretable, data-efficient decision support for Type 1 diabetes.

Skills

Education

Career History